Running DeepSeek R1 Locally on Windows with Ollama
· 12 min read · Views --
Last updated on

Running DeepSeek R1 Locally on Windows with Ollama

Author: Alex Xiang

Artificial Intelligence

DeepSeek R1 became very popular quickly, so I tried running it locally on Windows. The test machine was a Windows 11 laptop with an NVIDIA RTX 4060 Laptop GPU, 64GB memory, and 8GB VRAM.

This note records the basic setup: install Ollama, pull a DeepSeek model, test it in the terminal, and connect it to several user interfaces.

Install Ollama

On Windows, the simplest path is to download the installer from the Ollama website and run it directly.

After installation, open a terminal and check the version:

ollama --version

In my test, the version was:

ollama version is 0.5.7

Run DeepSeek R1

Ollama can pull and run the model in one command:

ollama run deepseek-r1:7b

The first run downloads the model files. After that, the model can be used directly from the terminal.

For machines with limited VRAM, the 7B model is a reasonable starting point. If performance is still too slow, a smaller model may be more practical.

One thing to note: reasoning models may output a visible thinking section depending on runtime and template behavior. That can make simple answers look verbose. For casual local use, this is acceptable, but for application integration you may want to control the output format more carefully.

Useful Ollama Commands

Common commands include:

ollama list
ollama run deepseek-r1:7b
ollama stop deepseek-r1:7b
ollama rm deepseek-r1:7b
ollama show deepseek-r1:7b

For local experimentation, ollama list and ollama ps are especially useful because they show which models are installed and which models are currently loaded.

Chatbox

Chatbox is a desktop AI client that can connect to local Ollama models. After installing Chatbox, choose the local model option, select Ollama as the API provider, and set the model to deepseek-r1:7b.

This is more comfortable than terminal interaction for everyday prompting. It is also a quick way to test whether the local Ollama API is working.

MaxKB

MaxKB is an open-source knowledge-base and RAG-style application. It can connect to Ollama and use a local model as its backend.

When MaxKB runs in Docker on Windows, remember that the container cannot access the Windows host through localhost. In many Docker Desktop setups, the correct host name is:

host.docker.internal

So the Ollama API address may need to be:

http://host.docker.internal:11434

After adding the model, MaxKB can be used to create a simple local knowledge-base application.

Open WebUI

Open WebUI is another good interface for local models. It is more focused on model management and chat experience. A typical Python installation flow is:

conda create -n open-webui python=3.11
conda activate open-webui
pip install open-webui
open-webui serve

After the service starts, open the local web UI in a browser and configure the admin account. If Ollama is running, Open WebUI should be able to list local models.

VS Code Integration

VS Code extensions such as Cline and Continue can connect to local or OpenAI-compatible model APIs. This makes it possible to use a local DeepSeek model for code explanation, small edits, and experimentation.

For serious coding tasks, local models still need careful review. The practical use case is not “let the model finish everything.” It is “use the local model as a cheap assistant for small, private, or repetitive tasks.”

Monitor GPU Usage

On NVIDIA machines, nvidia-smi is the first tool to check:

nvidia-smi -l 5

This refreshes GPU status every five seconds. Watch VRAM usage, GPU utilization, and running processes while the model is generating.

Final Takeaway

Running DeepSeek R1 locally on Windows is not difficult anymore. Ollama handles the model runtime, and tools like Chatbox, MaxKB, Open WebUI, and VS Code extensions make the model easier to use.

The main limitation is hardware. An 8GB GPU can run smaller models, but you should keep expectations realistic. Local models are excellent for experiments, private workflows, and low-cost daily assistance. For high-stakes reasoning or complex engineering tasks, stronger hosted models and human review are still valuable.